Predictive breeding in maize during the last 90 years

This article traces the 90-yr history and speculates on future applications of predictive breeding in maize (Zea mays L.). Predictive breeding started in the 1930s when superior double-cross maize hybrids were identified based on the mean of the four nonparental single crosses. The advent of recurrent selection in the 1940s led to methods to predict the mean of the next cycle of selection. The shift to single-cross hybrids in the 1960s necessitated methods to predict their performance, and genomic best linear unbiased prediction (GBLUP) was developed in 1994 for predicting single-cross performance for yield and other agronomic traits. In the 1990s, rapid recurrent selection with molecular markers led to the use of multiple regression for predicting the performance of individual plants undergoing selection. After Meuwissen, Hayes, and Goddard published their landmark article on genomewide selection in 2001, prediction methods shifted from multiple regression with fixed marker effects to ridge regression and Bayesian models with random marker effects. Subsequent research showed that GBLUP and ridge regression are equivalent when a trait is controlled by many small-effect loci distributed throughout the genome. Two trends that have strongly persisted in maize predictive breeding in the last 90 yr are leverage of resources and extraction of information from relatives. Predictive breeding in the 2020s will continue to expand in terms of what is being predicted and what predictive methods are used. Because predictions rely on having good phenotypic data for developing black-box prediction models, phenotyping will remain the cornerstone of predictive breeding.